feat(python): add native async sandbox clients

Signed-off-by: Max Dubrinsky <mdubrinsky@nvidia.com>

Author: maxdubrinsky

.gitignore

@@ -107,6 +107,7 @@ cython_debug/
 # Generated Python protobuf stubs (keep package marker)
 python/openshell/_proto/*
 !python/openshell/_proto/__init__.py
+python/openshell/aio.py
 
 # =============================================================================
 # IDE / Editor

docs/sdk/python.mdx

@@ -148,6 +148,45 @@ workspace for repeated exec, stop, start, and delete operations. List methods
 return lazy `Pager` instances; use `list_all()` only when you want to fetch the
 complete collection.
 
+## Use the Async Client
+
+`AsyncSandboxClient` provides the same sandbox operations and arguments as
+`SandboxClient` over `grpc.aio`. Construct and use it on one event loop:
+
+```python
+from openshell import AsyncSandboxClient
+
+async def run():
+    async with AsyncSandboxClient.from_active_cluster() as client:
+        sandbox = await client.get("worker", workspace="default")
+        result = await client.exec(sandbox.name, ["echo", "hello"], workspace="default")
+        print(result.stdout)
+
+        async for page in client.list(workspace="default"):
+            for item in page.items:
+                print(item.name)
+```
+
+List methods return lazy `AsyncPager` objects without an `await`. Use
+`await client.list_all(...)` to collect every page. The provider, sandbox-template,
+and workspace clients also have `Async` counterparts. Authentication settings
+match the synchronous client; credential refresh runs off the event loop.
+
+Both sandbox clients accept a caller-configured channel through `from_channel()`.
+That factory borrows the channel: closing the client leaves it open. Clients
+constructed from an endpoint or registered gateway own and close their channels.
+
+RPC methods accept `timeout` in seconds to override the client default. This is
+separate from an exec command's `timeout_seconds`. Mutations that support gateway
+deduplication accept `request_id`; reuse the same ID and payload when explicitly
+retrying an uncertain result. Request IDs do not enable general automatic retries. Exec retains the gateway
+owner-route fallback when the route fails before the first response; it never
+replays a command after receiving output.
+
+Async task cancellation propagates to the active RPC. For early exit from
+`exec_stream()`, use `contextlib.aclosing()` so the stream is cancelled promptly.
+Gateway RPC failures raise `GatewayError`, preserving the gRPC status and details.
+
 ## Call Generated RPCs
 
 The generated gRPC stubs include concrete request and response types for synchronous `grpc.Channel` and asynchronous `grpc.aio.Channel` connections. Install the dependency type stubs in your development environment for editor signatures and type checking:
@@ -177,7 +216,7 @@ async def get_provider(channel: grpc.aio.Channel, name: str) -> pb.ProviderRespo
     )
 ```
 
-Your application configures the channel's TLS and authentication, supplies RPC deadlines, and closes the channel. Create and use async channels on the same event loop. Cancellation and gRPC errors propagate directly. Generated stubs do not reuse `SandboxClient` authentication automatically; `SandboxClient` and `ProviderClient` remain synchronous.
+Your application configures the channel's TLS and authentication, supplies RPC deadlines, and closes the channel. Create and use async channels on the same event loop. Cancellation and gRPC errors propagate directly. Generated stubs do not reuse SDK client authentication automatically. Use `AsyncSandboxClient` for the higher-level async methods and managed authentication.
 
 If you annotate a stored async stub explicitly, use a deferred annotation such as `"openshell_pb2_grpc.OpenShellAsyncStub"`. That type exists only in the `.pyi` file; construct `OpenShellStub` at runtime. The same convention applies to credential-driver stubs.
 

pyproject.toml

@@ -78,7 +78,7 @@ git_describe_command = ["git", "describe", "--dirty", "--tags", "--long", "--mat
 [tool.ruff]
 target-version = "py311"
 line-length = 88
-exclude = ["python/openshell/_proto"]
+exclude = ["python/openshell/_proto", "python/openshell/aio.py"]
 
 [tool.ruff.lint]
 select = [

python/README.md

@@ -0,0 +1,18 @@
+# Python SDK implementation
+
+The synchronous clients in `openshell/sandbox.py` and `openshell/providers.py`
+own request construction and response conversion. `tasks/scripts/generate_python_async.py`
+emits concrete async counterparts into `openshell/aio.py`; the handwritten
+`openshell/_async_runtime.py` owns asyncio transport, authentication, pagination,
+and resource cleanup. Generated files are build outputs and are not committed.
+
+Run `mise run python:proto` after changing client methods or protobufs. The same
+generation runs when building release wheels. The async transform recognizes
+stub calls, lazy pager fetches, stream iteration, and polling sleeps. Add new I/O
+shapes explicitly to the generator and test their runtime and consumer typing
+contracts. Keep blocking operations out of generated async method bodies.
+
+Validate changes with Python formatting, lint, and type checks plus the affected
+tests. `python/openshell/async_client_test.py` exercises real loopback gRPC calls;
+`python/grpc_typing_test.py` checks consumer signatures. The wheel verifier requires
+both generated bindings and the async client module.

python/grpc_typing_test.py

@@ -79,3 +79,26 @@ def _check_types(path: Path) -> subprocess.CompletedProcess[str]:
         text=True,
         check=False,
     )
+
+
+def test_async_sdk_consumer_types(tmp_path: Path) -> None:
+    source = """
+from typing import assert_type
+from openshell import AsyncSandboxClient, AsyncPager, SandboxRef, ExecResult, DeletionResult
+
+async def check(client: AsyncSandboxClient) -> None:
+    assert_type(await client.get("worker", workspace="team"), SandboxRef)
+    assert_type(await client.create(workspace="team"), SandboxRef)
+    assert_type(client.list(workspace="team"), AsyncPager[SandboxRef])
+    assert_type(await client.list_all(workspace="team"), list[SandboxRef])
+    assert_type(await client.exec("worker", ["echo"], workspace="team"), ExecResult)
+    assert_type(await client.delete("worker", workspace="team"), DeletionResult)
+"""
+    path = tmp_path / "sdk_consumer.py"
+    path.write_text(source)
+    result = _check_types(path)
+    assert result.returncode == 0, result.stdout + result.stderr
+    path.write_text(source.replace('workspace="team"', "workspace=123"))
+    result = _check_types(path)
+    assert result.returncode != 0
+    assert "invalid-argument-type" in result.stdout

python/openshell/__init__.py

@@ -5,6 +5,14 @@
 
 from __future__ import annotations
 
+from ._async_runtime import AsyncPager
+from .aio import (
+    AsyncProviderClient,
+    AsyncSandboxClient,
+    AsyncSandboxSession,
+    AsyncSandboxTemplateClient,
+    AsyncWorkspaceClient,
+)
 from .errors import ErrorInfo, FieldViolation, GatewayError, from_grpc_error
 from .mutations import DeletionOutcome, DeletionResult
 from .providers import ProviderClient, ProviderCredentialValue
@@ -38,6 +46,12 @@
     __version__ = "0.0.0"
 
 __all__ = [
+    "AsyncPager",
+    "AsyncProviderClient",
+    "AsyncSandboxClient",
+    "AsyncSandboxSession",
+    "AsyncSandboxTemplateClient",
+    "AsyncWorkspaceClient",
     "ClientCredentialsAuth",
     "DeletionOutcome",
     "DeletionResult",